Instruments

An instrument is a capability class, not a vendor's model name — defined by benchmark band, context window and modality. You buy the class; the router picks the machine. This is what makes prices comparable and contracts durable when a new model ships.

On the live book today, model is the instrument / model id the matcher uses for a per-model order book. See How the spot market works.

GET /v1/markets is the machine-readable catalog used by Trade. It groups concrete models and capability instruments into families, preserves unknown seller model ids under Other, and includes live quotes for every configured region.

MODEL STRING	CLASS	CTX	MID $/M
`gpt-class-frontier`	Frontier chat & agents	200K	2.1800
`gpt-class-mid`	General production chat	128K	0.4204
`open-weights-70b`	Open weights, large	128K	0.0840
`open-weights-8b`	Routing, classification	32K	0.0191
`vision-long`	Document & video	1M	1.6400
`reasoning-xl`	Long-horizon reasoning	200K	6.9000
`embed-small`	Retrieval, rerank	8K	0.0110

Conformance

Each venue runs a rolling eval suite for its class. Drift beyond tolerance delists it from the book within minutes, without touching your code.

Pinning

If you must have one specific model, pass venue_pin. You'll pay closer to list — pinning removes the competition that makes the discount.

Regions and market data

There is still one price-time-priority book per model. A region is an eligibility filter over that global book, not a second regional matcher. Subscribe to /v1/orderbook?model_id=…&region=… for a filtered live view.

Historical charts come from GET /v1/markets/{model_id}/candles?region=…&interval=5m. Supported intervals are 1m, 5m, 1h, and 1d.

GET /v1/markets/{model_id}/estimate combines current compatible depth with a seven-day sample of sell-initiated executions. Its ETA and 24-hour suggested limit price are guidance, never a fill guarantee; sparse markets return an unavailable confidence level instead of invented precision.